Every developer knows the feeling. You reach for an AI agent to write a function, debug a test, or even outline a simple API endpoint. The code works. The problem gets solved. But a quiet concern gnaws at you: if the internet went down tomorrow, would you still be able to build the same thing from scratch? Atrophy, an open-source CLI tool created by developer Ashutosh Rath, exists to answer that question with a number and a chart.
Atrophy has accumulated 115 GitHub stars since its launch on July 5, 2026, and it is trending among developers who sense that their unaided skills are eroding but lack the data to prove it. The tool establishes baseline ratings for five cognitive skills, then tracks decay curves as you delegate more work to AI. It is MIT-licensed, written in TypeScript, and installable with a single npm command. The thesis is provocative: the same AI tools making developers productive are quietly making them dependent, and the only way to know for sure is to measure.
How Atrophy Works: Baselines, Drills, and the Unaided-vs-AI Chart
Atrophy runs entirely on your local machine with no telemetry and no cloud dependency. The workflow is straightforward. You start with atrophy baseline, a one-time session of roughly 25 minutes where you solve one exercise for each of five skills with AI assistance turned completely off. This establishes your starting Elo-style rating for each skill. From there, you run atrophy drill two or three times a week, spending 5 to 10 minutes on a single exercise automatically selected from the skill you have neglected the longest.
The five skills Atrophy measures cover the full range of how developers think. Syntax recall tests your ability to write a function from a spec against hidden tests. Debugging presents code with a single planted bug for you to find and fix. Code reading asks you to trace through a snippet and type exactly what it outputs. API memory tests your recall of standard library calls with blanked-out function names. Decomposition presents a design problem like building a rate limiter or folder sync tool and asks you to outline your approach, which you self-score against a revealed rubric.
Each drill is graded automatically. Passing raises your rating; failing drops it. The difficulty adapts to your skill level, selecting exercises where your predicted success rate lands near 65 percent. This is the sweet spot for deliberate practice where a single rep carries the most information. Comfortable wins teach the rating nothing, and impossible failures discourage progress.
The command atrophy serve launches a local dashboard at http://127.0.0.1:4646 that displays five skill curves with confidence bands and the chart that makes this entire tool worthwhile: unaided scores plotted in blue versus monthly AI-assisted scores plotted in green. If the blue line sinks while the green line stays perfect, that growing gap is your dependence, measured directly.
The Research Behind Atrophy: This Is Not Hypothetical
Atrophy is not just a clever idea. The README cites peer-reviewed evidence showing the pattern it measures is real and well-documented across professions. A 2025 study in The Lancet Gastroenterology and Hepatology found that doctors' unaided polyp-detection rates fell from 28 percent to 22 percent within months of routine AI assistance. A 2025 paper in PNAS showed that students who routinely used GPT-4 scored 17 percent worse than peers once the AI was taken away. A study of chess players found that cheating with chess engines created a permanent skill debt that did not recover even after the engines were removed.
The pattern has a name in the research literature: cognitive offloading. When a task is reliably automated, the brain stops maintaining the neural pathways required to perform it unaided. The insidious part is that the decline is invisible to the person experiencing it. People consistently feel fine while measurably declining because the AI fills the gap before they ever notice the gap exists. Atrophy makes the invisible visible by requiring periodic unaided reps and graphing the trend.
For AI-dependent developers, the implications are career-level. Job interviews for senior engineering roles still test unaided problem solving. Outages and air-gapped environments still happen. And the most effective engineers using AI tools are those who can evaluate the AI's output critically rather than accepting it wholesale. An Atrophy score trending downward is a leading indicator that you are building dependency rather than leverage.
Installation and Quick Start
Getting started with Atrophy requires Node.js 22 or later and takes under two minutes. The tool publishes to npm as atrophy and is available at version 0.4.0 as of this writing, still in active development with six open-source contributors and regular releases.
Install it globally: npm install -g atrophy. Then run your baseline: atrophy baseline. This walks you through one exercise per skill with AI off and takes about 25 minutes. After that, run drills regularly: atrophy drill. Each drill takes 5 to 10 minutes. The tool tracks your schedule and reminds you when you are overdue. To visualize your progress: atrophy serve opens the dashboard at port 4646 where you can see your ratings, confidence bands, and the unaided-versus-AI gap chart.
The exercise bank ships with hand-written problems in Python and JavaScript across three difficulty tiers, plus generator families that render endless fresh variants with randomized data, names, and twists. Your editor opens automatically via the $EDITOR environment variable, and grading runs your code against hidden tests in a sandboxed subprocess. Going over the soft time limit gradually reduces your score, but nothing explodes. You can keep fixing and resubmitting with the clock still running.
Who This Is For
Atrophy is for any developer who suspects their AI reliance has crossed the line from leverage into dependency but lacks the data to confirm it. It is for engineering teams that want to track skill health as a team metric the same way they track deployment frequency and incident response time. It is for solo founders who cannot afford to lose the unaided skills they will need when their AI tooling stack changes or goes down.
The tool is also relevant for engineering managers evaluating candidates. A candidate who maintains a high Atrophy rating while also shipping with AI assistance demonstrates the ability to use AI as leverage rather than crutch. That is a signal worth paying attention to. For the rest of us, Atrophy provides something we did not know we needed until we saw the chart: a honest look at whether the tools that make us faster are also making us weaker.
